Repeatable insight at scale

The title is a borrow (modified) from Patrick McKenzie.

My favorite frameworks (investment or otherwise) for generating insights have similar characteristics. They are:

• Fractal: Applicable to challenges of different sizes
• Transmittable: Can be passed on to another person in a relatively short period to support collaboration
• Stateless: Each use of the framework is not informed by previous uses

Fractal

The fractal characteristic is my attempt to describe how some frameworks provide insight on issues or challenges of different sizes. Useful frameworks utilize facets of the issues that are common irrelevant of the issue’s size. One of my issues with macro vs micro economic models is that they are intentionally non-fractal. The issue being examined has to fit primarily within the arbitrary macroeconomic or macroeconomic boundaries.

Fractal is also visual!

Source

Conversely, fluid dynamic models are a great example of modelling systems that provide insights at very small scale or very large scales. These models have become fashionable in modelling lots of systems with multiple individual agents.

Transmittable

Another characteristic is the ability to transmit (or communicate) the core principles of a framework to other people. Complexity isn’t always the reason that a framework is difficult to communicate but often the framework elements do not easily fit together or the outcomes of the framework are not repeatable. Another transmission issue might be that the framework involves multiple theories from different disciplines. I see this sort of Frankenstein’s monster approach in stock picking forums where people fuse theories of behavioral economics, macroeconomics, wisdom of crowds, statistics and mathematics together (think Fibonacci number analysis with ‘head and shoulders’ technical analysis plus ‘greed or fear’ crowd analysis as well as a nod to margin of safety valuation). They then tell people that its simple and easy to use!

A good framework has a central theory or premise, as all models and frameworks are merely mental structures to simplify the complexity of reality, and avoids multiple premises and assumptions. The results should provide predictive insight and the same information should generate the same approximate insights. I view incentives analysis, leaning on the evidence of human biases such as loss aversion, anchoring and optimism bias, as more easily transmittable and repeatable.

Stateless

The final characteristic is stateless which is borrowed from a software development concept about information retention. Stateful systems retain information on all previous uses or transactions whereas stateless systems do not retain any information of previous uses.

A stateless framework can be used anyone with little understanding of how it was previously used and there is no need for inputs from its previous use. It is a self-contained generator of insight and it is possible to start at day zero at any time.

Frameworks that rely on a detailed understanding of previous uses include neural networks. The opaque nature of neural networks means that I have to understand what data was used to train the model. I see the value of neural networks in discovering patterns in data that we haven’t seen before but I am hesitant to deploy neural networks as a strong framework for predictive insight, mainly because it is so previous information dependent.

In machine learning, I prefer cluster mining or k-means analysis where the data is clustered by different factors. This highlights different relationships but is not a set of weights or parameters that have been previously developed through training data to generate a particular answer.